Anatomy as a safety net: segmentation-aware checks for CT AI

Where false positives come from

On a target-negative cohort, Atlas flagged 13.8 percent of patients with at least one finding they did not have. Many false positives are anatomically implausible: an AAA call where the aortic caliber is normal, or an appendicitis call far from the right iliac fossa. A model that knows where organs are can catch some of these before they reach a radiologist.

Two layers of anatomic grounding

Atlas already grounds findings in anatomy at the readout level, mapping each detection to the nine-region abdominal grid that structured reporting uses [1][2]. The next layer is organ-level: validating detections against segmented structures so that a finding has to sit in a plausible place for its class. Robust whole-body CT segmentation across 104 structures is now available off the shelf [3], which makes this kind of plausibility check practical rather than aspirational.

Anatomy that spans regions

Anatomy also explains some of the model's limits. Acute pancreatitis frequently extends beyond the pancreas along fascial planes, so a single region label undercounts its true extent. That is not a model error so much as a property of the disease, recognized in the revised Atlanta classification [4][5], and it is a reason to treat region coverage as a nuanced metric rather than a pass/fail.